[diffusion] fix: fix accuracy for some image models

[mickqian]

Motivation

Modifications

1. Qwen-Image: align CFG with official diffusers by introducing true_cfg_scale, enabling CFG when tru e_cfg_scale > 1, and matching the official true-CFG norm rescale.
2. Qwen-Image-Edit: respect negative-image, so the uncond branch really uses negative_prompt.
3. Qwen-Image-Edit: keep Qwen2.5-VL vision rotary frequencies in fp32(aligned with diffusers) to reduce encoder drift
4. Qwen-Image-Edit with ulysses-degree=2: shard rope of condition-image together with rope of noisy-image, and build zero_cond_t modulation indices from local SP sequence lengths.
5. Z-Image: fix prompt/tokenization and dtype by using rendered chat-template prompts, bf16 text encoding, and fp32 latent sampling/scheduler state.
6. Z-Image with SP/Ulysses: stop sharding caption tokens (condition shouldn't be sharded), keep caption as a replicated suffix in joint attention, and fix the local/global RoPE offsets so multi-GPU stays aligned with single-GPU

Repro

Qwen-Image

$ sglang generate \
  --model-path Qwen/Qwen-Image \
  --attention-backend torch_sdpa \
  --prompt 'A coffee shop entrance features a chalkboard sign reading "Qwen Coffee  $2 per cup," with a neon
light beside it displaying "通义千问". Next to it hangs a poster showing a beautiful Chinese woman, and beneath
the poster is written "π≈3.1415926-53589793-23846264-33832795-02384197". Ultra HD, 4K, cinematic composition,
Ultra HD, 4K, cinematic composition.' \
  --negative-prompt " " \
  --width 1664 \
  --height 928 \
  --num-inference-steps 50 \
  --true-cfg-scale 4.0 \
  --seed 42 \
  --save-output

prompt = """A coffee shop entrance features a chalkboard sign reading "Qwen Coffee  $2 per cup," with a neon light beside it displaying "通义千问". Next to it hangs a poster showing a beautiful Chinese woman, and beneath the poster is
written "π≈3.1415926-53589793-23846264-33832795-02384197". Ultra HD, 4K, cinematic composition"""

pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16).to("cuda")
image = pipe(
    prompt=prompt,
    negative_prompt=" ",
    width=1664,
    height=928,
    num_inference_steps=50,
    true_cfg_scale=4.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]
  image.save("/tmp/triplet_outputs/qwen_image_diffusers.png")

before this branch	this branch	diffusers

with --ulysses-degree 2:

Qwen-Image-Edit

$ sglang generate --model-path=Qwen/Qwen-Image-Edit-2511 --prompt="Convert 2D style to 3D style" --image-path="https://github.com/lm-sys/lm-sys.github.io/releases/download/test/TI2I_Qwen_Image_Edit_Input.jpg" --save-output --width=1536 --height=1024

before this branch	this branch	diffusers

abs_diff (1-gpu vs diffusers):

with --ulysses-degree 2:

Z-Image-Turbo

$ sglang generate \
  --model-path Tongyi-MAI/Z-Image-Turbo \
  --prompt 'Young Chinese woman in red Hanfu, intricate embroidery. Impeccable makeup, red floral forehead pattern. Elaborate high bun, golden phoenix headdress, red flowers, beads. Holds round folding fan with lady, trees, bird. Neon lightning-bolt lamp (⚡️), bright yellow glow, above extended left palm. Soft-lit outdoor night background, silhouetted tiered pagoda (西安大雁塔), blurred colorful distant lights.' \
  --seed 42 \
  --height 1024 \
  --width 1024 \
  --num-inference-steps 9 \
  --guidance-scale 0.0

import torch
from diffusers import ZImagePipeline

prompt = """Young Chinese woman in red Hanfu, intricate embroidery. Impeccable makeup, red floral forehead pattern.
Elaborate high bun, golden phoenix headdress, red flowers, beads. Holds round folding fan with lady, trees, bird. Neon
lightning-bolt lamp (⚡️), bright yellow glow, above extended left palm. Soft-lit outdoor night background, silhouetted
tiered pagoda (西安大雁塔), blurred colorful distant lights."""

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=False,
).to("cuda")
image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    num_inference_steps=9,
    guidance_scale=0.0,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("/tmp/triplet_outputs/zimage_turbo_diffusers.png")

before this branch	this branch	diffusers

with --ulysses-degree=2:

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses a specific issue within the qwen-image model's implementation, ensuring that sequence prefixes are correctly managed during the forward pass. This fix is essential for the model's stability and accurate processing of multimodal inputs, particularly in scenarios involving text sequence lengths.

Highlights

• Qwen-Image Model Fix: Corrected the forward method in the qwen_image.py file by adding the num_replicated_prefix argument, set to seq_len_txt, to ensure proper handling of sequence prefixes for multimodal processing.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

• python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py
  • Added num_replicated_prefix argument to a function call in the forward method.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

The pull request correctly introduces the num_replicated_prefix parameter to the USPAttention call, using seq_len_txt. This ensures that replicated text tokens are handled appropriately within the attention mechanism, which is crucial for maintaining correct attention weights in a distributed setup. The change aligns with the intended functionality of the USPAttention module.

[mickqian]

/tag-and-rerun-ci

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[mickqian]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[yhyang201]

/rerun-failed-ci

[Rockdu]

Thanks for fixing this SP precision issue, here are some quantized test results on our side for reference

Worst-case min_cosine vs single_gpu_ref

Model	Metric	Before	After	Result
Z-Image-Turbo	model_output	0.5039	1.0000	✅ Fixed (exact match)
Z-Image-Turbo	prev_sample_mean	0.7694	1.0000	✅ Fixed (exact match)
Qwen-Image	model_output	0.8616	1.0000	✅ Fixed (exact match)
Qwen-Image	prev_sample_mean	0.8825	1.0000	✅ Fixed (exact match)